Starling claims first bank to train AI on UK sovereign cloud

Starling Bank says it is the first bank to train an artificial intelligence model on a UK sovereign cloud. The bank plans to use Civo’s UK infrastructure to train a transactional foundation model with Nvidia on more than 15 billion anonymised transactions and other events.

Training on real data is due to begin after several months of development and testing with synthetic data. Starling expects the model to enter production in early 2027.

The first planned use is fraud detection, followed by Starling Assistant, the bank’s AI financial assistant. The model is designed to identify patterns in sequences of transactions and related events. It will not process language in the same way as a large language model.

Sebastian Khan, Starling’s data scientist and technical lead for the project, described the model as Starling’s own system, built from the ground up. He said it is designed to “crunch and understand transactions”.

A regular coffee purchase could have a different meaning if it happened during a holiday, Khan explained. The price and type of purchase might be similar, but its location and the transactions around it could change.

The model will use tokenised data rather than personally identifiable information. Starling will replace raw values, such as postcodes, with coded numbers represented in a mathematical vector space. The bank will keep the system that maps the original values to the codes separately from the training data.

Tokenisation will take place in Starling’s existing Google Cloud environment, which contains its data warehouse. The encoded data will then be transferred to Civo’s UK cloud infrastructure. Khan described the arrangement as creating no additional data sovereignty risk.

Civo chief executive Mark Boost described the project as part of Starling’s multicloud strategy. Starling is using a British cloud provider, while Civo is UK-owned and operates datacentres in the UK.

The sovereignty claim covers the location and control of the cloud infrastructure. Nvidia will provide the graphics processing hardware and much of the software used to train the model. Nvidia’s CUDA software ecosystem is updated from the United States.

Boost acknowledged that a fully UK-based supply chain would be preferable, but the highest-performing AI chips are currently made by US companies. He identified control of the system as the relevant issue, asking whether an external provider could switch it off. “These are silicon chips that sit in a datacentre in the UK, and Nvidia can’t turn it off as such,” he said.

The model is expected to be trained initially on more than 15 billion events, although Starling described that number as preliminary. Khan said the system could improve fraud detection by capturing the order and context of transactions, reducing reliance on manually created indicators.

The project is linked to the UK government’s plan to encourage AI adoption in financial services. The plan identifies the sector’s reliance on a small number of global cloud and AI providers as an issue.

Starling has not decided how often the model will be retrained. The schedule will depend partly on when the training data and the model’s predictions become outdated.